Meeting News

AI system successfully identifies patients with FH

Daniel J. Rader

ATLANTA — A machine learning model was able to analyze large health-care encounter databases to identify patients with familial hypercholesterolemia, according to data presented at the FH Foundation’s FH Global Summit.

Findings from this study were simultaneously published in Lancet Digital Health.

“It’s fair to say that we think based on all of this data, precision screening for FH is now a reality,” Daniel J. Rader, MD, chair of the department of genetics, chief of the division of translational medicine and human genetics and the Seymour-Gray Professor of Molecular Medicine at the University of Pennsylvania Perelman School of Medicine, said during the presentation. “This markedly reduces the number of people to screen to actually find individuals with FH, making this concept of precision screening in a population-based level using these big data, both claims and lab data as well as electronic health record data a reality.”

Kelly D. Myers

“What this study shows is we’ve been able to develop a cutting-edge tool that leverages machine learning and big data to address the No. 1 problem with familial hypercholesterolemia, and that being lack of diagnosis,” Kelly D. Myers, chief technology officer for the FH Foundation, told Cardiology Today. “The model and the process make precision screening a reality by reducing the number of individuals who need to be screened dramatically. If you are hypothetically hoping to identify eight individuals with FH, you would need to screen about 2,000 individuals in a population-based screening approach.”

Machine learning model

Researchers developed the machine learning model with electronic health record data from 939 patients who were clinically diagnosed with FH and 83,136 patients who were free from FH from four U.S. institutions. Data used in the development of this machine learning model included procedure, prescription, laboratory and diagnosis information.

A machine learning model was able to analyze large health-care encounter databases to identify patients with familial hypercholesterolemia, according to data presented at the FH Foundation’s FH Global Summit.
Source: Healio/Cardiology Today

Once developed, the machine learning model was then validated and tested using data from a large tertiary care academic medical center (n = 173,733) and a national health care encounter database (n = 170,416,201).

All patients from the development, validation and testing cohorts had at least one CVD risk factor including hypercholesterolemia, hypertension or hyperlipidemia.

“A key issue that I think many people don’t appreciate in this work is that individuals with a previous FH diagnosis were excluded,” Rader said during the presentation. “The idea is to identify people who don’t already have a diagnosis of FH.”

In the national database, patients with values above the threshold for probable FH were flagged. From there, the FH Foundation contacted the patient’s physician to see whether they wanted to know the identity of the identified patients from their practice. If the patient’s identity was revealed, the doctor would then determine the likelihood of FH, which was then categorized as definite, probable, possible, inconclusive or unlikely. Patients who were considered to have possible or greater likeliness of FH were sent for further screening and follow-up. A similar process was used for validation in the large tertiary care academic medical center.

The machine learning model had a positive predictive value of 0.85, a sensitivity of 0.45, an area under the receiver operating characteristic curve of 0.89 and an area under the precision-recall curve of 0.55.

The model was able to flag 866 patients from the health care delivery system dataset and 1,331,759 patients from the national database as likely to FH. Of these patients, FH experts reviewed a sample of 45 patients from the national health care dataset and 103 patients from the health care delivery system dataset.

Of the patients from the sample reviewed by FH experts, 77% of those in the health care delivery system dataset (95% CI, 68-86) and 87% of patients from the national database (95% CI, 73-100) were considered to have high enough clinical suspicion of FH to justify the initiation of guideline-based clinical evaluation and treatment.

Potential implications

Katherine Wilemon

“Physicians only have so much time with a patient, especially in the United States with a health care system that is so fragmented,” Katherine Wilemon, founder and CEO of the FH Foundation, told Cardiology Today. “One of the interesting things that the application of machine learning can do in an algorithm model like FIND FH is to look across an individual’s entire medical history and assess at what age did this person first have, for example, in this case, high cholesterol mark. It may not have even been noted. ... What that means for other conditions is very exciting because physicians will never have enough time to review this much data on their own and see those patterns in an individual or across a population.”

In a related editorial published in Lancet Digital Health, Alexandre C. Pereira, MD, PhD, professor of genetics and molecular medicine at the University of Sao Paulo Medical School and director of the human genetics group of the Laboratory of Genetics and Molecular Cardiology at the Heart Institute in Sao Paulo, wrote, “Hopefully providers and payers will embrace the challenge and start to use already available technology to bring into fruition the cost savings that large-scale identification of familial hypercholesterolemia can deliver, These implementations will pave the way to truly transformative care of individuals with familial hypercholesterolemia.”

Additional research using this machine learning model is already underway, Rader said.

“A study that we’ve been working on now for a couple years in collaboration with the [FH] Foundation is IN-TANDEM,” he said. “The objective of this study is to establish a relationship between a machine learning to find an FH score and the likelihood of having either an FH mutation or an FH clinical diagnosis in two of the main hospitals within the Penn system.” – by Darlene Dobkowski

References:

Rader DJ. Session II: Precision Targeting and Screening to Find FH. Presented at: The FH Foundation’s FH Global Summit; Oct. 20-21, 2019; Atlanta.

Myers KD, et al. Lancet Digit Health. 2019;doi:10.1016/S2589-7500(19)30150-5.

Pereira AC. Lancet Digit Health. 2019;doi:10.1016/S2589-7500(19)30161-X.

Disclosures: The study was funded by the FH Foundation and supported by Amgen, Regeneron and Sanofi. Rader reports he serves on an advisory board for Alnylam, Novartis and Pfizer and is an unpaid advisor to the FH Foundation. Wilemon is an employee of the FH Foundation. Myers is a paid consultant for the FH Foundation. Pereira reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.

Daniel J. Rader

ATLANTA — A machine learning model was able to analyze large health-care encounter databases to identify patients with familial hypercholesterolemia, according to data presented at the FH Foundation’s FH Global Summit.

Findings from this study were simultaneously published in Lancet Digital Health.

“It’s fair to say that we think based on all of this data, precision screening for FH is now a reality,” Daniel J. Rader, MD, chair of the department of genetics, chief of the division of translational medicine and human genetics and the Seymour-Gray Professor of Molecular Medicine at the University of Pennsylvania Perelman School of Medicine, said during the presentation. “This markedly reduces the number of people to screen to actually find individuals with FH, making this concept of precision screening in a population-based level using these big data, both claims and lab data as well as electronic health record data a reality.”

Kelly D. Myers

“What this study shows is we’ve been able to develop a cutting-edge tool that leverages machine learning and big data to address the No. 1 problem with familial hypercholesterolemia, and that being lack of diagnosis,” Kelly D. Myers, chief technology officer for the FH Foundation, told Cardiology Today. “The model and the process make precision screening a reality by reducing the number of individuals who need to be screened dramatically. If you are hypothetically hoping to identify eight individuals with FH, you would need to screen about 2,000 individuals in a population-based screening approach.”

Machine learning model

Researchers developed the machine learning model with electronic health record data from 939 patients who were clinically diagnosed with FH and 83,136 patients who were free from FH from four U.S. institutions. Data used in the development of this machine learning model included procedure, prescription, laboratory and diagnosis information.

A machine learning model was able to analyze large health-care encounter databases to identify patients with familial hypercholesterolemia, according to data presented at the FH Foundation’s FH Global Summit.
Source: Healio/Cardiology Today

Once developed, the machine learning model was then validated and tested using data from a large tertiary care academic medical center (n = 173,733) and a national health care encounter database (n = 170,416,201).

All patients from the development, validation and testing cohorts had at least one CVD risk factor including hypercholesterolemia, hypertension or hyperlipidemia.

“A key issue that I think many people don’t appreciate in this work is that individuals with a previous FH diagnosis were excluded,” Rader said during the presentation. “The idea is to identify people who don’t already have a diagnosis of FH.”

PAGE BREAK

In the national database, patients with values above the threshold for probable FH were flagged. From there, the FH Foundation contacted the patient’s physician to see whether they wanted to know the identity of the identified patients from their practice. If the patient’s identity was revealed, the doctor would then determine the likelihood of FH, which was then categorized as definite, probable, possible, inconclusive or unlikely. Patients who were considered to have possible or greater likeliness of FH were sent for further screening and follow-up. A similar process was used for validation in the large tertiary care academic medical center.

The machine learning model had a positive predictive value of 0.85, a sensitivity of 0.45, an area under the receiver operating characteristic curve of 0.89 and an area under the precision-recall curve of 0.55.

The model was able to flag 866 patients from the health care delivery system dataset and 1,331,759 patients from the national database as likely to FH. Of these patients, FH experts reviewed a sample of 45 patients from the national health care dataset and 103 patients from the health care delivery system dataset.

Of the patients from the sample reviewed by FH experts, 77% of those in the health care delivery system dataset (95% CI, 68-86) and 87% of patients from the national database (95% CI, 73-100) were considered to have high enough clinical suspicion of FH to justify the initiation of guideline-based clinical evaluation and treatment.

Potential implications

Katherine Wilemon

“Physicians only have so much time with a patient, especially in the United States with a health care system that is so fragmented,” Katherine Wilemon, founder and CEO of the FH Foundation, told Cardiology Today. “One of the interesting things that the application of machine learning can do in an algorithm model like FIND FH is to look across an individual’s entire medical history and assess at what age did this person first have, for example, in this case, high cholesterol mark. It may not have even been noted. ... What that means for other conditions is very exciting because physicians will never have enough time to review this much data on their own and see those patterns in an individual or across a population.”

In a related editorial published in Lancet Digital Health, Alexandre C. Pereira, MD, PhD, professor of genetics and molecular medicine at the University of Sao Paulo Medical School and director of the human genetics group of the Laboratory of Genetics and Molecular Cardiology at the Heart Institute in Sao Paulo, wrote, “Hopefully providers and payers will embrace the challenge and start to use already available technology to bring into fruition the cost savings that large-scale identification of familial hypercholesterolemia can deliver, These implementations will pave the way to truly transformative care of individuals with familial hypercholesterolemia.”

PAGE BREAK

Additional research using this machine learning model is already underway, Rader said.

“A study that we’ve been working on now for a couple years in collaboration with the [FH] Foundation is IN-TANDEM,” he said. “The objective of this study is to establish a relationship between a machine learning to find an FH score and the likelihood of having either an FH mutation or an FH clinical diagnosis in two of the main hospitals within the Penn system.” – by Darlene Dobkowski

References:

Rader DJ. Session II: Precision Targeting and Screening to Find FH. Presented at: The FH Foundation’s FH Global Summit; Oct. 20-21, 2019; Atlanta.

Myers KD, et al. Lancet Digit Health. 2019;doi:10.1016/S2589-7500(19)30150-5.

Pereira AC. Lancet Digit Health. 2019;doi:10.1016/S2589-7500(19)30161-X.

Disclosures: The study was funded by the FH Foundation and supported by Amgen, Regeneron and Sanofi. Rader reports he serves on an advisory board for Alnylam, Novartis and Pfizer and is an unpaid advisor to the FH Foundation. Wilemon is an employee of the FH Foundation. Myers is a paid consultant for the FH Foundation. Pereira reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.

    See more from FH Global Summit